Nvidia’s Kepler to use different CUDA structure

According to 3D Center Nvidia are planning a radically different architecture on Kepler. The design will retain the same basic hierarchy present on the Fermi cards (400/500 series) but will employ a different implementation of the CUDA layout. The first cards expected are going to be based off the GK104 architecture which has its specifications listed below, this will include the anticipated GTX 680. Interestingly both the HD 7970 and GTX 680 will possess 32 ROPs and 128 TMUs while in terms of Graphics cores Nvidia’s card has 1536 CUDA cores and the HD 7970 has 2048 stream processors. The GTX 680 is also only expected to pack 2GBs of GDDR5 instead of the 3 on the HD 7970, whilst the HD 7970 also contains a larger memory bus at 384bits instead of 256. Whether any of this will make a difference we are yet to see since Nvidia are claiming a performance win but the specifications seem to show it will be close.

SIMD Hierarchy

• 4 Graphics Processing Clusters (GPC)
• 4 Streaming Multiprocessors (SM) per GPC = 16 SM
• 96 Stream Processors (SP) per SM = 1536 CUDA cores

TMU / Geometry Domain

• 8 Texture Units (TMU) per SM = 128 TMUs
• 32 Raster OPeration Units (ROPs)

Memory

• 256-bit wide GDDR5 memory interface
• 2048 MB (2 GB) memory amount standard

Clocks/Other

• 950 MHz core/CUDA core (no hot-clocks)
• 1250 MHz actual (5.00 GHz effective) memory, 160 GB/s memory bandwidth
• 2.9 TFLOP/s single-precision floating point compute power
• 486 GFLOP/s double-precision floating point compute power
• Estimated die-area 340mm²

Source: 3DCenter.org